BIO-AI: STAR: Machine Learning for Robust Demographic Inference Under Biologically Realistic Conditions

Across the tree of life, populations diverge upon isolation by geographic barriers, exchange migrants upon secondary contact, and adapt to environmental pressures. These processes leave signatures in species’ genomes, which can be used to understand the factors shaping biodiversity. However, popular methods for disentangling these signatures are limited both in terms of efficiency and accuracy, and Artificial Intelligence (specifically, machine learning) offers a powerful alternative. Despite recent advances, machine learning approaches have yet to reach their potential in this field and remain limited in the processes they can consider, their applicability across organisms, and their accessibility to researchers with varying levels of technical expertise. The proposed work will develop robust, user-friendly machine learning tools for investigators studying the drivers of diversification. Furthermore, the proposed work will use these tools to illuminate the evolutionary histories of several empirical systems, including fruit flies, mosquitoes, plants, snails, and slugs. By creating well-documented, user-friendly tools, this work will provide a valuable resource to the broader community of evolutionary biologists. Furthermore, the work will support NSF’s desired societal outcome of the development of a globally competitive workforce by hosting workshops (both virtual and in-person), and training a postdoctoral researcher, a graduate student, several undergraduates, and high school students in machine learning and software development. The overall objectives of the proposed work are to develop robust, user-friendly machine learning tools for population genetic inference and to apply these tools to uncover the drivers of diversification in several empirical systems. The rationale for the proposed work is that it will facilitate more accurate assessment of the processes driving diversification across the tree of life based on genomic data. The proposed work will accomplish these overall objectives by pursuing three specific aims. (1) Expand the current implementation of popai, a Python package for demographic inference using machine learning, and compare it to state-of-the-art methods. This aim will expand the diversity of data formats, network architectures, evolutionary models, and inferential tasks available in popai. (2) Develop machine learning tools to detect model violations and make robust inferences in their presence. The proposed work will explore the impacts of several model violations likely to be prevalent in empirical systems and develop new tools for detecting violations. Further, the proposed work will use domain adaptation, a machine learning approach, to arrive at robust conclusions in the presence of model violations. (3) Apply these tools to uncover the drivers of diversification in several empirical systems, including a co-distributed group of terrestrial invertebrates and plants from the Pacific Northwest of North America. The proposed work is novel and important because it will allow researchers to specifically identify those model violations likely to mislead inference and to arrive at accurate inferences in the presence of such model violations without directly modeling them. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria. NSF Award ID: 2552066 | Program: 01002627DB NSF RESEARCH & RELATED ACTIVIT | Principal Investigator: Megan Smith | Institution: Mississippi State University, MISSISSIPPI STATE, MS | Award Amount: $399,999 View on NSF Award Search: https://www.nsf.gov/awardsearch/show-award/?AWD_ID=2552066 View on Research.gov: https://www.research.gov/awardapi-service/v1/awards/2552066.html